Try Int32(0) + kv_coord += 1 (matching reference pattern)

This commit is contained in:
2026-05-22 21:59:53 +00:00
parent 7e832aa527
commit 8c9d4eb1ef

View File

@@ -217,10 +217,12 @@ class FmhaV3StageCMulti:
cute.copy(tma_q, tAgQ[(None, Int32(0))], tAsQ[(None, qh.index)], tma_bar_ptr=qh.barrier)
qp.tail()
kvp.reset()
kv_coord = Int32(0)
for kt in cutlass.range(0, n_kv_tiles, 1, unroll=1):
kvh = kvp.acquire_and_advance()
cute.copy(tma_k, tBgK[(None, kvh.count)], tBsK[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
cute.copy(tma_v, tVgV[(None, kvh.count)], tVsV[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
cute.copy(tma_k, tBgK[(None, kv_coord)], tBsK[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
cute.copy(tma_v, tVgV[(None, kv_coord)], tVsV[(None, kvh.index)], tma_bar_ptr=kvh.barrier)
kv_coord += 1
kvp.tail()
# ===== MMA warp =====